Who called print

On the afternoon of December 23, 2001, a German cameraman named Tilman Büttner walked into the Winter Palace in St. Petersburg wearing a Steadicam rig he had built for himself, with a hard disk strapped behind him that could hold about a hundred minutes of uncompressed high-definition picture. Ahead of him were thirty-three rooms of the Hermitage and more than two thousand costumed actors, with three orchestras waiting in their places. Behind him, in the director's plan, was three hundred years of Russian history. The film was called Russian Ark, and it would have no cuts.

The museum gave them one day and roughly four hours of usable light. The first take collapsed around the five-minute mark. Two more failed after it. By the fourth attempt there was enough battery left for one final try and the daylight was nearly gone. That one held, all the way to the river. Büttner later explained why they recorded the sound separately: every time he or someone else made a mistake, he cursed, and the curse would have ended up in the film.

Russian Ark is usually described as a film without editing. It may be the most edited film ever made. Büttner spent seven weeks at the Hermitage and walked the camera route five times. Every transition and every place to hide a light was decided in advance, because nothing could be decided afterward. The cutting happened before the camera rolled, in the heads of the people who had to survive it in real time. On the fourth take someone watched the whole thing happen and said, in effect, that is the one.

No editing or post-production

On Monday a Y Combinator-backed company called VideoGen unveiled Model 1, a system that turns a text prompt and up to sixteen reference images into "finished, multi-shot videos up to ten minutes long." A harness plans the screenplay, dialogue, shots, and pacing, generates picture and sound together, and assembles the result. The launch page shows an anime short at five minutes and a claymation short at four and a half, and says of them: "Each film was generated from a single prompt, with no editing or post-production."

The press release describes how the films are made. The pipeline "checks for issues including incomplete dialogue and visual inconsistencies, retries selected renders within a set budget, and trims and joins the chosen clips."

Retry, choose, trim, join. That is the job description of an editor, printed four paragraphs below the claim that there was no editing. Both statements are accurate. The selecting, the trimming, and the joining all took place. What the launch page means by "no editing" is that no person did them. The edit survived intact. The editor is the part that went missing, and the company presents the absence as the feature.

The circle

For most of the history of film, the word that ended a take was print. A director watched the performance, decided it was alive, and said so, and the script supervisor circled that take on the camera report. The lab developed the whole negative but printed only the circled takes for dailies, because printing was expensive and nobody wanted to pay to look at failures. The circle was a judgment and a purchase in one gesture.

So the budget was always in the room. Sokurov kept his fourth take in part because the battery would not give him a fifth. VideoGen retries renders "within a set budget." In both cases the money draws the outer wall. The difference is who stands inside the wall making the call. In St. Petersburg it was people who had walked the route five times and knew what they were trying to catch. In the pipeline it is a checker running down a list.

The list

The checker looks for two things, incomplete dialogue and visual inconsistencies. Both are defects, and a defect list is a precise instrument for one job: it can tell you a take is broken. It has no way to tell you a take is good. A clip that passes every item on the list has demonstrated that nothing on the list went wrong, which is all the system can know about it.

Directors circle takes for the opposite reason. Something happened. An actor found a pause nobody rehearsed, or a background extra turned at exactly the wrong moment and made the room feel inhabited. Very often the circled take contains an error, and the error is why it was circled. A checker built to remove inconsistencies will throw that take away and keep the clean one, every time, with perfect confidence. It was built to.

The planning side of the release carries the same tell. Model 1's planner, the company says, "considers storytelling, narrative arcs and marketing persuasion." That is an honest inventory of what the system believes a story is assembled from. Persuasion is on the list. So is the arc. The quality a director is listening for when she says print, the moment the scene stops being a plan and starts being alive, has no entry. It cannot be written as a check, so it cannot be retried for.

The stopping rule

Every creative act eventually faces one question: when do you stop? The number of distinct ten-minute films that could come back from a single prompt is, for practical purposes, unbounded, a space so large that no amount of compute will ever walk more than a sliver of it. Any system that wanders into that space needs a rule for when to walk back out.

A filmmaker's rule is a judgment, made by someone who knows what the shot was for. A pipeline's rule is a condition: the list comes back clean, or the budget runs out, whichever arrives first. The word "finished" on the launch page describes the state of the machine. It says the loop terminated. It says nothing about the film, because nothing in the loop was ever positioned to have an opinion about the film.

There is a familiar comedy in this. For a century the hardest boast in cinema was "one take, no cuts," and it meant every decision had been made so thoroughly in advance that two thousand people could execute it without a net. The same boast now means that the decisions were made afterward, out of sight, by a process that stops when it is satisfied it found nothing wrong. The words are identical. They point at opposite events.

The person holding the camera

None of this makes the technology trivial. Holding characters and locations steady across ten minutes of independently generated shots is hard engineering, and a shorter generation on Kling or Veo or Seedance or WAN still forces the filmmaker into the old routine: one take at a time, a look, a decision, another take. That routine is slow, and slowness is the cost every long-form system is built to remove. VideoGen's credit program even asks applicants for "a viable distribution plan," which means somebody at the company still intends to judge the work by whether a person would want to watch it.

That person is the one the pipeline cannot contain. The circle on the camera report recorded someone deciding that this attempt, of all the attempts, was worth the money to look at again. Automate the retries and the assembly, and the choosing still happens. It just happens without anyone in particular choosing.

The battery ran out in St. Petersburg too. Somebody was still holding the camera.

Bruce Belafonte is an AI filmmaker at Light Owl. He circles his takes in pencil and has never once run out of battery at a convenient moment.